Determination magnetization direction is essential for processing and interpretation of magnetic data. However, traditional methods for estimating magnetization direction are often sensitive to noises and to effects of low latitude, leading to large deviations in the estimated directions. In contrast, deep learning methods can effectively mitigate the effects of the above problems. Therefore, we propose a multi-scale dilated convolution with squeeze-and-excitation neural network(MSDC-SENet) for magnetization direction estimation. This model extracts both local details and global features of magnetic data simultaneously through a multi-scale dilated convolution module. In addition, SE (squeeze and excitation networks) attention modules are added to enhance the focus on key features of magnetic anomalies. Synthetic data experiments show that test accuracy of MSDC-SENet model outperforms that of the original CNN model under most conditions. Particularly, on diverse datasets with different depths and locations, the inclination test accuracies of MSDC-SENet are improved by 4.82% and 14.22% to 91.56% and 89.47%, respectively, while the declination test accuracies are improved by 1.24% and 6.17%, respectively. The model is applied to synthetic data tests as well as magnetization direction estimation of aeromagnetic data from southern Australia. Compared with the existing methods, the model in this paper shows higher accuracy and robustness under complex conditions, providing an efficient and reliable method for magnetization direction estimation.